多策略改进河马算法的BiGRU滚动轴承寿命预测

Multi-strategy hippopotamus algorithm for rolling bearing life prediction by BiGRU

关键词:  剩余寿命预测  河马优化算法  双向门控循环单元  健康曲线  滚动轴承

Key words:  remaining life prediction  hippopotamus optimization algorithm  bi-directional gated recurrent unit  health curve  rolling bearing

摘要:

为更加精准地预测滚动轴承剩余寿命,提出一种基于多策略改进河马优化(TOBCHO)算法与双向门控循环单元(BiGRU)融合的轴承寿命预测方法。首先,对全寿命周期信号进行特征提取,通过构建基于相关性、单调性和鲁棒性综合评价指标,筛选出敏感特征向量作为敏感特征集,并采用主成分分析(PCA)技术构建健康指标曲线。其次,针对BiGRU在滚动轴承寿命预测研究超参数难以确定的问题,提出多策略改进河马算法优化BiGRU的寿命预测模型(TOBCHO-BiGRU),该模型在河马算法种群初始化阶段引入最优最差对立学习机制,生成对立解扩展搜索空间；采用混沌映射序列替代传统随机数的生成方式以解决算法收敛不稳定的问题；在HO迭代后期引入最优个体自适应分布扰动策略,通过动态调整扰动强度以实现局部开发与全局搜索能力的平衡。最后,通过在国际通用的IEEE PHM 2012轴承数据集上进行实验验证,并与其他多种预测模型进行对比分析,结果充分证明了所提TOBCHO-BiGRU方法在预测精度方面具有显著优势。消融实验结果表明,各改进策略之间存在正向协同效应,共同促进河马优化算法性能的提升,为复杂工况下的滚动轴承寿命预测提供了高精度解决方案。

Abstract:

To predict the remaining life of rolling bearings more accurately, this paper proposed a bearing life prediction method based on the fusion of the multi-strategy hippopotamus algorithm (TOBCHO: adaptive t-distribution, optimal-worst opposite learning, and chaos mapping) and bi-directional gated recurrent unit (BiGRU). Firstly, feature extraction was performed for the whole life cycle signal, and comprehensive evaluation indexes were established based on correlation, monotonicity, and robustness. Sensitive feature vectors were screened as the sensitive feature set, and principal component analysis (PCA) technology was used to construct the health index curve. Then, for the problem that it was difficult for BiGRU to determine the hyperparameters in rolling bearing life prediction research, a life prediction model of TOBCHO-optimized BiGRU (TOBCHO-BiGRU) was proposed, which introduced the optimal worst opposition-based learning mechanism in the population initialization stage of the hippopotamus algorithm and generated the opposing solution to expand the search space. The chaos mapping sequence was adopted to replace the generation of random numbers to solve the problem of unstable convergence of the algorithm. The adaptive distribution perturbation strategy for optimal individual was introduced in the late iteration stage of the hippopotamus optimization (HO) algorithm, and the perturbation strength was dynamically adjusted to balance the local development and global search capability. Finally, the experimental validation was conducted on the internationally used IEEE PHM 2012 bearing dataset, and the proposed model was compared with a variety of other prediction models. The results adequately show that the proposed TOBCHO-BiGRU method has a significant advantage in terms of prediction accuracy. Ablation experiment results demonstrate that there are positive synergistic effects among improvement strategies, promoting the enhancement of the HO algorithm, which provides a high-precision solution for the rolling bearing life prediction under complex working conditions.